Run Details
The run page (/run/[id]) shows up to 10 tabs. Tabs appear only when the run has data for them.
| Tab | Content |
|---|---|
| Config | Hyperparameters, env, git snapshot |
| Metrics | All scalar curves — line charts with log axis, smoothing, Step / Wall Time toggle |
| Histograms | Weight distributions over steps (three-pane evolution) |
| PCA | 3D scatter of embeddings (Three.js), with step slider and auto-play |
| Landscape | 2D loss-landscape grids — heatmap, contour overlay, 3D surface; step slider + auto-play |
| Figures | Logged images / matplotlib / G2 figures |
| Texts | Logged text samples |
| Media | Audio / video players |
| Tables | Logged tabular data |
| Model | PyTorch model graph (mviz) |
Metrics tab
Section titled “Metrics tab”- Each metric card supports moving-average smoothing, log axis, and multi-series overlay
- X-axis can toggle between Step and Wall Time
- Hover shows crosshair + tooltip
Landscape tab
Section titled “Landscape tab”Each logged landscape (t.log_loss_landscape(...), see Advanced Data Types) becomes a card:
- Heat — pure colormap field, hover reads (α, β, loss)
- Contour — pure contour lines (no fill)
- Both — heatmap + d3-contour overlay
- 3D — interactive Three.js surface mesh: drag to rotate, scroll to zoom, wireframe toggle, Front/Side/Top views
- ⚽ Roll + speed — replays a gradient-descent ball (from the highest point, with trail) in every view; speed selectable 0.5×–4×; auto-plays on view switch and on each new step frame
- Colormap picker — coolwarm (default) / plasma / magma / viridis, applied to heatmap, contours and the 3D mesh
- Z scale (log / lin) — default log: an offset-log scale expands near-minimum detail so the bowl stays visible even when extreme walls dominate the min–max range; contour levels densify near the bottom accordingly; the 3D loss axis is labeled
(log). Hover tooltips always show the raw loss - Step slider + ▶ auto-play — replay landscape evolution across training (slider moves between logged frames — log per epoch for step-by-step playback)
- Cards group by name and refresh live while the run is still training
Sharing a run
Section titled “Sharing a run”The Share button generates an anonymous read-only link:
https://host/run/{id}?token=xxxAnyone with the link can view the run without logging in.